Paper: What a Design Tool for the Age of Agents Looks Like (#AI4SDLC)
I watched Y Combinator’s Design Review episode “How To Design In The Agent Era,” published on August 7, 2026. Stephen Haney, founder of the design tool Paper, joins YC partner Aaron Epstein to review websites submitted by viewers. It is ostensibly a redesign show, but I was more interested in something else: Paper is a rare tool designed for agents from the ground up rather than retrofitted for them.
First, some context. Haney has been working on the designer–developer handoff problem for years: he co-founded Modulz and helped create Radix UI, one of the most popular React component libraries. After Modulz was acquired by WorkOS, he spent two years leading product there. Paper is his second attempt to bridge the same gap, this time with agents in mind.
The engineering idea comes down to one architectural choice. Traditional design tools render their canvases with proprietary engines, so an agent has to work through exports and plugin APIs: translating between formats consumes tokens, adds latency, and creates a separate class of hallucinations. In Paper, the canvas is live HTML and CSS: every object on screen is the web code itself. An agent—Cursor, Claude Code, or another tool—connects through an MCP server and writes literal HTML, which the canvas renders without a translation layer. Haney’s formula is that a good handoff for people automatically becomes a good handoff for agents.
The workflow looks like this: a designer leaves comments or uses the Paper Snapshot extension to capture a live website as editable layers; an agent generates variants, even overnight; and a person curates the result by deleting, branching, and refining. According to Haney, their brand designer built Paper’s own marketing site in the tool and took it all the way to Next.js code with animations and shaders in about a week, without engineers.
The practical part of the episode is a catalogue of familiar AI-generated design tells: bold type everywhere, five to eight font sizes instead of three, ALL-CAPS subheads with wide letter spacing, pill badges on every card, purple gradients—the training-data skew that followed Linear’s success—card-heavy sections, and a dark-mode toggle in an MVP. Haney describes the source as an insecure designer baked into the model weights, and his remedy is subtraction: much of design is deleting. The same pattern recurs in the live redesigns of Legion Health, Sytex, and Moreta: the copy and value proposition are usually fine, while the visual composition and hierarchy fall apart.
It is also revealing how the Paper team uses agents itself: 12 people work with Cursor, Claude Code, and Bugbot, but humans review every line of code. The canvas has to sustain consistently high performance, and agents still cannot deliver that level of system quality on their own. Haney’s principle is simple: use agents to accelerate the people.
This connects neatly with things I have covered before. WebMCP applies the same idea to the web: instead of asking an agent to “look at the pixels and guess,” the website exposes explicit tools. Geoffrey Litt’s talk argues that generation is getting cheaper while understanding becomes the bottleneck; Haney makes the same point about taste. Agents quickly learn tactics such as font weights, contrast, and spacing, but choosing a direction and discovering the “next aesthetic” remain human work. And the story of a Meta product manager who launches products while barely knowing how to code shows the same role shift for designers: less manual production, more framing and curation.
My main takeaway is this: artifacts designed only for human eyes become a tax on the agent stack. Tools win when human and machine access share one contract—a design the agent can read and edit as code, and code the designer can see as a canvas. Whether Paper truly is “the fastest-growing design tool since Figma” remains a claim from YC and the company itself, but the direction looks right to me.
P.S. After watching the video, I decided I should try the tool in practice :)
#AI4SDLC #AI #Agents #DevTools #Product #Engineering